Linchenn

Thank you shanumantesc for the detailed investigation! Let me check this.

Sorry for the delay. I will check it soon and will share you the updates when I have.

@shanumantesc Could you provide the codes that returns wrong results? I just tried the following codes but the results are correct: ```javascript const a = tf.tensor3d([1, 2, 3, 4, 5,...

From my understanding, the shader compiler here uses tensor's logic shape (un-squeezed shape): https://github.com/tensorflow/tfjs/blob/8c7fd026bb9940c926a94f70d7bee5ef1f51a1ef/tfjs-backend-webgl/src/shader_compiler.ts#L1058 As the result, the cache key construction here are also supposed to use tensor's logic shape...

@shanumantesc Sorry, I could not reproduce this:  Did I miss something?

Thank you Yang! I have reached a consensus over the new model list with Na, Ping and Jiajia (I have pinged you on the spreadsheet), so I will update the...

Thank you Jiajia @qjia7! This is a great inspiration. Let me see if we can reuse this algorithm and, at least, we could have a packed version of `Conv2DDerInputProgram` to...

(The tool currently does not work for WebGL https://honry.github.io/webnn-samples/style_transfer/?backend=webgl, `Binary was compiled with different shapes than the current args. Shapes 1,8,270,270 and 1,270,270,8 must match`, so I manually tested `conv2d`...

Do you have examples/hints for other algorithms? @qjia7 As we known, WebGL's `conv2d` has three ways (1*1, img2col and brute force). Do you mean a new way to compute such...

Great catch! Let me check if I could benchmark im2col and matmul separately first. Thank you!